Silicon Language: A Robot-Native Knowledge Exchange Framework for Heterogeneous Robots

Yi Liu; Xianglin Meng; Chang Chen; Jingjing Fan

cs.ROOct 6, 2026

Silicon Language: A Robot-Native Knowledge Exchange Framework for Heterogeneous Robots

Authors: Yi Liu, Xianglin Meng, Chang Chen, Jingjing Fan

Organizations: School of Mechanical Engineering, Beijing Institute of Technology, Beijing, China · Yulin Saiyi Intelligent Technology Co., Ltd., Yulin, China · Yulin Intelligent Unmanned Equipment Innovation Center Co., Ltd., Yulin, China

Abstract

Reusing a capability across heterogeneous robots still requires substantial human adaptation and verification: transferring a skill often means re-engineering interfaces, retuning parameters, and re-validating safety. We introduce Silicon Language, a robot-native knowledge exchange framework that treats the robot as the active subject of its own capability evolution. In this framework, a robot that wants a capability encodes its own experience into knowledge packets, publishes them, retrieves peer packets, translates them for its own sensors and actuators, and reviews them through independent local trial. A receiver-side usability evaluation procedure lets each robot decide for itself whether an external packet is useful, and progressive blending with automatic rollback is designed to reduce the risk of negative transfer when adopting it. The system combines three infrastructure layers (edge agent, Silicon Transfer Protocol (STP), and knowledge hub) with a capability stack inspired by the human scholarly system. We report a 30-day proof-of-concept deployment at an above-ground simulated-mine laboratory in Yulin, with following trials on an outdoor sand road and an indoor factory floor. Two heterogeneous robots encoded and translated three capabilities across embodiments through operator-assisted file copies mediated by the Silicon Language translation layer; source-side trials of the dust-locked following behavior were recorded on Taurus. Project records indicate that per-capability adaptation time dropped from days to hours; we present these figures as descriptive deployment records rather than controlled measurements. The deployment provides initial evidence for cross-embodiment knowledge exchange; fleet-level autonomous evolution and controlled with/without-packet comparisons remain future work.

Figures & tables

Fig. 1: Silicon Language reference architecture. The three-layer infrastructure (edge agent, STP, knowledge hub) supports a seven-layer capability stack inspired by the human scholarly system. The stack is a functional organization, not a claim of equivalence with human scholarship.

Human system	Function	Silicon Language layer
Writing	Solidify experience	L7 Writing
Publishing	Enter shared repository	L6 Publishing
Citation / search	Build on prior work	L5 Retrieval
Translation	Cross-domain reuse	L4 Translation
Peer review	Validate claims	L3 Review
Versioning	Continuous revision	L2 Versioning

TABLE I: Mapping between the human scholarly system and the Silicon Language capability stack.

Field	Content
Packet ID	kp.behavior.following.dust_locked.v1
Producer	Taurus ( guide_vehicle_tracking package, Autoware-era branch, $\sim$ 2026-Q2)
Type	Code knowledge (processing-logic gene: 9-stage pipeline + state machine)
Version	1.0.0 (supersedes: none)
Body (pipeline)	S1 ROI crop $\rightarrow$ S2 ground removal $\rightarrow$ S3 voxel downsample $\rightarrow$ S4 Euclidean cluster $\rightarrow$ S5 OBB fit $\rightarrow$ S6 target selection $\rightarrow$ S7 state machine (IDLE/SEARCHING/LOCKED/REACQUIRE/SAFE_STOP) $\rightarrow$ S8 distance-to-speed lookup $\rightarrow$ S9 lateral controller (Ackermann $\delta$ or differential $\omega$ )
Body (key knobs)	Distance-to-speed breakpoints $\{0,2,3.5,5,7,9,12,15\}$ m $\mapsto$ speeds $\{0,0,1.0,1.6,2.5,4.0,6.0,6.0\}$ m/s; stop distance 5 m; miss threshold 20 frames; reacquire timeout 3 s

TABLE II: Example knowledge packet for dust-locked lead-vehicle following, produced on Taurus and adapted for Di-Ting. The body represents processing logic, while implementation code and message bindings depend on the robot embodiment.

Fig. 2: Structure of a Silicon Language knowledge packet. The header specifies identity and applicability; the body supports five knowledge forms; metadata records training, environment, hardware, and performance information.

Panel of Figure 5 — Fig. 3: Receiver-side usability evaluation. The receiver pre-screens, translates, trials, and verdicts each packet, with progressive blending and automatic rollback mitigating negative-transfer risk.

Message	Role	Layer
KNOWLEDGE_ANNOUNCE	Advertise a new packet	L6 Publishing
KNOWLEDGE_REQUEST	Query the hub for packets	L5 Retrieval
KNOWLEDGE_OFFER	Propose a packet in response	L5 Retrieval
GENE_STREAM	Fragmented packet transfer	Transport
ACKNOWLEDGMENT	Receipt and agreement	L1 Dialogue
FITNESS_REPORT	Local review verdict	L3 Review

TABLE III: STP message types and their corresponding capability-stack layers. The protocol is specified but not yet deployed in production; the table describes the reference design.

Panel of Figure 7 — Fig. 4: The two robots that exchanged knowledge through Silicon Language during the deployment. Top: both platforms in the mine test area—foreground Di-Ting (belt inspection robot) and background Taurus (transport robot) on a test ramp. Bottom left: front view of Di-Ting, showing the camera/lidar mast, front lights, and communication antennas. Bottom right: Taurus, showing the larger transport deck, rail guides, and roof-mounted lidar/radar. The size, sensor mounting, and locomotion differences are the embodiment gaps the framework translates across.

	Di-Ting (inspection)	Taurus (transport)
Primary role	Belt/tunnel inspection	Material transport
Chassis type	Differential drive	Ackermann steering
Dimensions (m)	$0.94\times 0.83\times 0.84$	$2.01\times 1.03\times 1.00$
Wheelbase/track (m)	track 0.66	wheelbase 1.09
Mass	200 kg	500 kg
Payload	50 kg	300 kg

TABLE IV: Platform differences relevant to cross-embodiment transfer. Values are from chassis configuration files and author-confirmed hardware facts.

Fig. 5: Longitudinal velocity measured on Taurus across the five recorded following trials. Trial 1 ( sand_road_real_track_1 ) is a 12.5 s stationary pre-roll with zero velocity; the remaining four trials contain motion. All velocities are chassis-feedback values from the ROS bags.

Fig. 6: Commanded velocity versus measured chassis velocity for trial 2, the longest recorded following run (273.7 s, 290.4 m). The figure illustrates the controller’s response and lag; it is not a lock-retention or tracking-error metric, because the bags do not record the behavior-layer target state.

Fig. 7: Peak, p95, and mean absolute speeds for each Taurus trial. Trial 1 is omitted because it recorded no motion. The behavior-layer max_speed was configured to 5.0 m/s; the peak observed speed (5.58 m/s in trial 5) slightly exceeds this configured limit because of CAN-feedback quantization on a downhill segment. Commanded velocity never exceeded 5.56 m/s, which remains below the Taurus chassis rated maximum of 9.72 m/s (35 km/h).

Trial	Duration (s)	Distance (m)	Peak (m/s)	Mean (m/s)	p95 (m/s)	Brakes	Outcome
1 sand_pre	12.50	0.0	0.0	0.0	0.0	0	stationary pre-test
2 sand_long	273.71	290.43	5.556	1.074	5.528	5	not labelled
3 sand_mid	242.03	140.00	3.583	0.585	2.222	4	not labelled
4 sand_short	230.88	184.01	3.861	0.802	3.056	1	failed (lost lead)
5 factory	278.30	157.51	5.583	1.710	4.472	4	not labelled

TABLE V: Descriptive motion metrics for the five recorded Taurus following trials. Distances are integrated from chassis velocity; outcomes are operator labels. These metrics are descriptive and do not by themselves attribute performance to the Silicon Language framework.

Fig. 8: Onboard RGB frames from Taurus trial 4 (640 $\times$ 480, YUV422). (a) $t=100$ s: following the lead vehicle. (b) $t=118$ s: the operator-labelled loss moment—the lead vehicle is off-center to the front-right. (c) $t=130$ s: robot stopped after human intervention. The frames are shown only as qualitative evidence of the failure event; they do not support quantitative claims about perception accuracy.

Fig. 9: Di-Ting velocity profile during the downhill human-track failure. After an initial peak near 0.65 m/s, the chassis speed collapses to essentially zero as the behavior enters SAFE_STOP; the gear report remains in drive (value 10), confirming that the stop is commanded by the behavior rather than by a gear shift.

ID	Packet	Trigger	Action	Status
RB-001	kp.behavior.following.intelligent.v1	Guider lost on downhill; recovery timeout reached	SAFE_STOP: brakes to $v=0$ , publishes STATUS_FAILED	observed (Di-Ting)
RB-002	kp.behavior.following.dust_locked.v1	miss_counter $\geq$ threshold during dust burst	SAFE_STOP path in state machine	coded, not observed

TABLE VI: Observed SAFE_STOP events and coded rollback paths. RB-001 is a source-side safety stop, not an adoption rollback; RB-002 is a hypothetical path present in the source code with no corroborating bag.

Fig. 10: Development calendar-time records for three capabilities (log scale). Slope detection by manual coding on the transport robot was not completed within the project budget and is marked as such rather than plotted with a proxy value. Times are calendar-time estimates from project records, not controlled measurements; modes are not like-for-like, so we report no aggregate speedup ratio.

Explore similar work

Oct 8, 2026cs.RO

RoboRSI: Stable, efficient, and reusable robot self-evolution in complex real-world environments

A generalist robot should not only perform diverse tasks but also improve through experience, turning what it learns during execution into capabilities that later tasks can reuse. Robot agents that act through code can already repair programs from execution feedback, yet it remains a central challenge to organize this experience around the task structure that gives it meaning, so that each repair is attributed to the responsible capability, supported by execution evidence, and validated before it is reused. We introduce RoboRSI, a robot self-improvement system built on Top-Down Skill Refinement (TSR). TSR decomposes tasks into compound, atomic, and base skills with scoped responsibilities and explicit input--output contracts, attributes each execution outcome to the responsible branch, and confines revision to that branch. Building upon this structure, a Manager, Planner, Engineer, and Reviewer coordinate planning, execution, diagnosis, and the validated release of new skills, while people steer the process through objectives and corrections; stable skill sequences are further consolidated into reusable compound skills. On a mobile manipulator, RoboRSI develops multi-object household cleanup over 104 rounds. In simulation, it achieves the highest success rate on LIBERO, LIBERO-PRO, LIBERO-Plus, and RoboTwin, exceeding the strongest baseline by 2.7 to 11.0 percentage points.

Zimo Wen, Yijin Chen, Yuxuan Cao +8

Topic pending

Ask PDF

Jun 30, 2026cs.RO

ASPIRE: Agentic /Skills Discovery for Robotics

Traditional robot programming is challenging: it requires orchestrating multimodal perception, managing physical contact dynamics, and handling diverse configurations and execution failures. We introduce ASPIRE (Agentic Skill Programming through Iterative Robot Exploration), a continual learning system that autonomously writes and refines robot control programs in a code-as-policy paradigm while compounding experience into a reusable skill library. ASPIRE discovers skills that persist across tasks, simulation and real-world settings, and embodiments. It operates in an open-ended loop with three components: (1) a closed-loop robot execution engine that exposes fine-grained multimodal traces, enabling autonomous failure diagnosis, repair synthesis, and validation; (2) a continually expanding skill library that distills validated fixes into reusable, transferable knowledge; and (3) evolutionary search that generates diverse task sequences and control programs to explore beyond single-trajectory refinement. ASPIRE surpasses prior methods by up to 77% on LIBERO-Pro manipulation under perturbation, 72% on Robosuite bimanual handover, and 32% on BEHAVIOR-1K long-horizon household tasks. Its accumulated library also enables zero-shot generalization to unseen long-horizon tasks: on LIBERO-Pro Long, ASPIRE achieves 31% success versus 4% for prior methods despite their use of test-time reasoning and retries. Finally, simulation-discovered skills provide initial evidence of sim-to-real transfer, substantially reducing real-robot programming effort across different embodiments and robot APIs.

Runyu Lu, Yubo Wu, Ethan Kou +11

NVIDIA · UMich · UIUC +2

Robot Policy Learning Long-Horizon Robotic Manipulation Continual Robot Learning Robot Skill Learning Generalization in Robotic Manipulation Program Synthesis+4

Ask PDF

Apr 22, 2026cs.RO

MOMO: A framework for seamless physical, verbal, and graphical robot skill learning and adaptation

Industrial robot applications require increasingly flexible systems that non-expert users can easily adapt for varying tasks and environments. However, different adaptations benefit from different interaction modalities. We present an interactive framework that enables robot skill adaptation through three complementary modalities: kinesthetic touch for precise spatial corrections, natural language for high-level semantic modifications, and a graphical web interface for visualizing geometric relations and trajectories, inspecting and adjusting parameters, and editing via-points by drag-and-drop. The framework integrates five components: energy-based human-intention detection, a tool-based LLM architecture (where the LLM selects and parameterizes predefined functions rather than generating code) for safe natural language adaptation, Kernelized Movement Primitives (KMPs) for motion encoding, probabilistic Virtual Fixtures for guided demonstration recording, and ergodic control for surface finishing. We demonstrate that this tool-based LLM architecture generalizes skill adaptation from KMPs to ergodic control, enabling voice-commanded surface finishing. Validation on a 7-DoF torque-controlled robot at the Automatica 2025 trade fair demonstrates the practical applicability of our approach in industrial settings.

Markus Knauer, Edoardo Fiorini, Maximilian Mühlbauer +10

German Aerospace Center (DLR), Institute of Robotics and Mechatronics (RMC), Münchener Str. 20, 82234 Weßling, Germany · School of Computation, Information and Technology (CIT), Technical University of Munich (TUM), Arcisstr. 21, 80333 Munich, Germany

Robot Skill Learning Robot Imitation Learning Industrial Robotics Language-Conditioned Robot Control Human-Robot Interaction Robot Learning

Ask PDF